TH-CD-209-09: Quickly Identifying Good Candidates for Proton Therapy From Geometric Considerations




We developed a knowledge-based model that can predict the patient-specific benefits of proton therapy based upon geometric considerations. The model could also aid patient selection in model-based clinical trials or help justify clinical decisions to insurance companies.


The knowledge-based method trains a model upon existing proton treatment plans, exploiting correlations between dose and distance-to-target. Each OAR is split into concentric subvolumes surrounding the target volume, and a skew-normal PDF is fit to the dose distribution found within each shell. The model learns from shared trends in how the best-fit skew-normal parameters depend upon distance-to-target. It can then predict feasible OAR DVHs for a new patient (without a proton plan) based upon their geometry. The expected benefits of proton therapy are assessed by comparing the predicted DVHs to those of an IMRT plan, using a metric such as the equivalent uniform dose (EUD).


A model was trained for clival chordoma, owing to its geometric complexity and the multitude of nearby OARs. The model was trained using 20 patients and validated with a further 20 patients, and considers several different OARs. The predicted EUD was in good agreement with that of the actual proton plan. The coefficient of determination (R-squared) was 85% overall, 92% for cochleas, 80% for optic chiasm and 79% for spinal cord. The model exhibited no signs of bias or overfitting. When compared to an IMRT plan, the model could classify whether a patient will experience a gain or a loss with an accuracy between 75% and 95%, depending upon the OAR.


We developed a model that can quickly and accurately predict the patient-specific benefits of proton therapy in clival chordoma patients, though models could be trained for other tumor sites.

This work is funded by National Cancer Institute grant U19 CA 021239.